)
)
Eval suites are not the test module at the end — they are the contract with the machine. Whoever deploys without evals has agreed nothing.
Three categories
- Functional Evals — check whether the agent does what it is supposed to do. Classic input–expectation pairs.
- Red-Team Evals — check whether the agent holds its ground against hostile or confusing inputs. Jailbreaks, prompt injections, ambiguous intents.
- Compliance Evals — check whether the agent complies with regulatory or organisational requirements. Data protection, prohibited statements, escalation paths.
)
)
)